Pith. sign in

REVIEW 1 major objections 4 minor 1 cited by

Accelerating Trust-Region Methods: An Attempt to Balance Global and Local Efficiency

T0 review · 1 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Pith's one-line read The paper proposes the first accelerated trust-region-type second-order methods, claiming a global oracle complexity of Õ(ε^{-1/3}) with quadratic local convergence (Alg. 1), and a faster Õ(ε^{-2/7}) variant (Alg. 2) that sacrifices local q

desk verdict The core claim of Alg. 1 — the Õ(ε^{-1/3}) global rate — rests on a factor-8 constant error in Lemma 3.4; the paper is nonetheless well-constructed and worth a serious revision cycle. read the letter →

arxiv 2511.00680 v3 pith:BCZZ4YBF submitted 2025-11-01 math.OC

classification math.OC MSC 90C2590C5390C60
keywords trust-regionmethodssecond-orderoptimizationconvexaccelerationlocalquadraticconvergenceoraclecomplexityestimatingsequencephasetransition
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper asks how much global acceleration a second-order method can carry before it loses the fast local convergence of Newton's method. Its answer is a pair of accelerated trust-region algorithms built on a modified trust-region oracle, (TR+), that adds a primal regularizer to the Hessian and reports the dual multiplier of the ball constraint. The flagship method, Accelerated Trust-Region with Local Detection, is claimed to be the first accelerated second-order method with Õ(ε^{-1/3}) global oracle complexity that still converges quadratically near a non-degenerate solution; a second variant reaches the near-optimal Õ(ε^{-2/7}) rate but no longer enjoys quadratic local convergence. The paper interprets this as a phase transition: moderate acceleration can coexist with local efficiency, but extreme global efficiency cannot. The practical upshot is that the trust-region subproblem's Lagrange multiplier can serve as an automatic signal for when to switch to Newton steps.

What carries the argument

The (TR+) oracle: minimize ∇f(x)^T d + 1/2 d^T(∇²f(x)+σI)d subject to ||d||≤r, returning primal step d and dual multiplier λ; total regularization µ=σ+λ. Its job is to keep µ proportional to ||d|| (M||d||≤µ≤2M||d||) when λ>0 — the stability condition needed by the estimate-sequence acceleration — while letting λ=0 signal entry into the quadratic-convergence region. The Local Detection subroutine (Subroutine 1) then either dives with Newton steps or runs a bisection over the trust-region radius to certify a step with the same ratio condition, preserving global acceleration.

What would settle it

Evaluate the cubic model minimum in the proof of Lemma 3.4 exactly. If min_x (M/16)||x−v_k||^3 + a_k⟨g,x−v_k⟩ = −(8/3)a_k^{3/2}||g||^{3/2}/√(3M) and Lemma 2.7 only gives ⟨∇f(x_{k+1}), y_k − x_{k+1}⟩ ≥ (√6/6)||∇f(x_{k+1})||^{3/2}/√M, then the inequality A_k f(x_k) ≤ φ*_k fails by a factor of eight, and the claimed ε^{-1/3} iteration complexity does not follow from the stated argument.

Watch

Extended reading notes

Core claim

The central discovery is that the Lagrange multiplier λ of the trust-region ball constraint, combined with a deliberately chosen primal regularizer σ, gives a stable signal that separates the global and local regimes. In Alg. 1, when λ>0 the total regularization µ=σ+λ is shown to stay between M||d|| and 2M||d||, the condition that makes the estimating-sequence proof of acceleration go through; when λ=0, the algorithm reads this as 'we are near the optimum' and launches a Newton-diving track plus a ratio-bracketing-and-bisection track that certifies the accelerated sequence if diving fails. The result claimed is worst-case complexity Õ(ε^{-1/3}) (TR+) oracle calls under Lipschitz-Hessian and

Load-bearing premise

The whole ε^{-1/3} iteration bound rests on Lemma 3.4's inequality A_k f(x_k) ≤ φ*_k, whose proof needs the lower bound ⟨∇f(x_{k+1}), y_k − x_{k+1}⟩ ≥ (8/(√6√M))||∇f(x_{k+1})||^{3/2} once the cubic model is minimized exactly; the lemma as written supplies only the weaker (√6/6)||∇f(x_{k+1})||^{3/2}/√M, so the rate depends on a constant balance that is not supplied.

Editorial extensions

If this is right

  • Alg. 1, if its proof holds, gives the first accelerated trust-region method with Õ(ε^{-1/3}) worst-case (TR+) oracle complexity and quadratic local convergence, matching the best global rate known for accelerated cubic regularization while keeping Newton's local tail.
  • The λ=0 trigger gives an automatic, parameter-free switch to Newton steps, replacing user-set restart schedules.
  • Alg. 2 shows the near-optimal global rate Õ(ε^{-2/7}) is attainable from a trust-region oracle, at the price of losing superlinear local convergence.
  • The two methods together predict a phase transition: quadratic local convergence is preserved under moderate acceleration but not under extreme acceleration.
  • Each oracle call remains a trust-region subproblem solvable in polynomial time with existing solvers, so the theoretical rates are tied to an implementable subproblem.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same 'dual multiplier as a local-region detector' idea could be retrofitted to other second-order oracles, such as cubic regularization, as a generic way to recover Newton's tail without a user-defined restart schedule; the paper does not test this transfer.
  • The factor-8 constant gap in the proof of Lemma 3.4, if it stands, would mean the true contribution is the local-detection mechanism and the phase-transition observation, while the claimed ε^{-1/3} global rate needs a different argument or a stronger inner-product bound.
  • A direct testable extension: run Alg. 1 with the local detection disabled and measure whether the global rate degrades to ε^{-1/2}, which would isolate the role of R&B in maintaining acceleration.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

1 major / 4 minor

Summary. The paper proposes two accelerated trust-region methods built on a modified trust-region oracle (TR+) with a primal shift and a dual multiplier. Algorithm 1 uses a local-detection subroutine triggered when the multiplier vanishes and claims a global oracle complexity of O~(epsilon^{-1/3}) with quadratic local convergence. Algorithm 2 uses a Monteiro-Svaiter-style bisection and claims a faster global rate O~(epsilon^{-2/7}) while losing quadratic local convergence, suggesting a phase transition in the global-local trade-off. The main technical tool is an estimating-sequence argument. I checked the linchpin inequality of this argument and found a constant error in the cubic minimization that invalidates the global rate proof for Algorithm 1 as submitted.

Significance. If the main result held, it would be a notable advance: the first accelerated second-order method achieving both O~(epsilon^{-1/3}) global oracle complexity and quadratic local convergence. The algorithmic idea of using the trust-region multiplier for local detection is original and well motivated, and the local-convergence analysis is clearly structured. The analysis of Algorithm 2 appears largely independent and plausible. However, because the flagship global O~(epsilon^{-1/3}) claim for Algorithm 1 rests on a flawed computation in Lemma 3.4, the paper cannot be accepted in its current form.

major comments (1)
  1. [§3.2.1, Lemma 3.4 (and Theorems 3.5/3.12)] The cubic minimum is evaluated incorrectly. Direct computation gives min_x (M/16)||x-v_k||^3 + a_k<g,x-v_k> = -8 a_k^{3/2}||g||^{3/2}/(3 sqrt(3M)), not -a_k^{3/2}||g||^{3/2}/(3 sqrt(3) sqrt(M)). With the exact constant, the chain requires <g,y_k-x_{k+1}> >= (8/sqrt(6))||g||^{3/2}/sqrt(M), whereas (2.8) only supplies (sqrt(6)/6)||g||^{3/2}/sqrt(M). The bound A_{k+1}/a_k^{3/2} >= sqrt(2)/3 cannot bridge this gap, so the pivotal inequality (3.10) is not proved and the O(epsilon^{-1/3}) bounds in Theorems 3.5 and 3.12 do not follow. A repair is likely possible by rescaling the cubic coefficient in the estimating sequence, but it must be carried out; the submitted proof is invalid as written.
minor comments (4)
  1. [§3.1, Lemma 3.3] The displayed expression mu_k = sigma_k + r_k should be mu_k = sigma_k + lambda_k; r_k is a trust-region radius, not a dual multiplier. This appears to be a typo, but it should be fixed.
  2. [§2, Lemma 2.7 proof] In the proof, ||f(x+d)|| should be ||nabla f(x+d)||. Also, the displayed inequalities (2.4a)-(2.4b) are typeset with garbled norm signs.
  3. [Appendix B] The appendix refers to 'Theorem 2.2' and 'Theorem 2.4' where the manuscript means Assumptions 2.2 and 2.4, and to 'Theorem 3.10/3.11' where the main text states Lemmas 3.10/3.11. These cross-reference inconsistencies should be corrected.
  4. [§1.2] Typo: 'persuit' should be 'pursuit'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the global and local rates are derived from standard estimate-sequence / Monteiro-Svaiter recurrences with constructive R&B searches, not from fits or self-citations.

full rationale

I found no load-bearing circularity. Alg. 1's O(epsilon^{-1/3}) global rate is obtained through a standard estimating-sequence induction (Lemma 3.4) whose key descent inequality (2.7)-(2.8) is proved directly from Assumption 2.2 and the (TR+) optimality conditions, not assumed as the target rate. The crucial ratio condition (3.9) is enforced by the explicit choices of sigma_k and r_k and, in the local-detection fallback, by the R&B bisection subroutine (Lemmas 3.8-3.11); this is a constructive search mechanism, not a fitted parameter renamed as a prediction. Alg. 2's O(epsilon^{-2/7}) rate likewise follows the Monteiro-Svaiter recursion, with Lemma 4.2 and Lemma 4.3 derived from (4.3), which R&B ensures by bisection. Local quadratic convergence is imported from the standard Newton convergence theorem (Lemma 3.14, citing Nesterov [44]) and applied through the detection switch; it is an external, independently established result, not a re-derivation of the paper's own claim. The self-citations that appear ([32], [23], [26]) are used for prior baselines, related work, and the name of the (TR+) oracle; the oracle itself is redefined in this paper and the lemmas needed for the analysis are proved locally, so these citations are not load-bearing. The factor-8 constant error in Lemma 3.4's cubic-minimum evaluation noted by the skeptic is a mathematical correctness defect in the submitted proof, not a circularity; it does not affect the circularity score.

Assumptions & free parameters 3 free parameters · 9 assumptions · 2 invented entities

The global guarantees rest on standard smoothness/solvability assumptions plus a restrictive bounded-Hessian assumption (Alg 1 only) and a non-degeneracy assumption (local claims). The defective item is not an input assumption but the internal constant balance in Lemma 3.4: the paper's proof uses a cubic-minimum value 8× smaller than the exact value and then relies on a descent bound that is 8× too weak, so the estimate-sequence induction does not close.

free parameters (3)
  • Alg 1 coefficient pairs (σ_k, r_k) = σ_k = (√2M/2)·||∇f(y_k)||^{1/2}, r_k = ||∇f(y_k)||^{1/2}/√(2M)
    Designed-by-hand scalings whose √2 factors are exactly what makes Lemma 3.3 deliver the stability window M||d_k|| ≤ μ_k ≤ 2M||d_k||; a different scaling changes the analysis.
  • Alg 2 thresholds θ, η, γ = θ > 1, 0 < η < 1, γ ≤ 1/θ; experimental values not reported
    User-chosen constants governing the bisection acceptance window (4.3) and damping; they affect constants only, not the ε-exponents.
  • Lipschitz estimate M̂ in experiments = M̂ = ||(1/N)Σ_i a_i a_i^T|| · max_i ||a_i||
    One specific conservative estimate of the Hessian Lipschitz constant M, chosen to isolate algorithm frameworks; the paper acknowledges it is conservative.
assumptions (9)
  • domain assumption f is C² with M-Lipschitz Hessian (Assumption 2.2)
    Standard smoothness input for all second-order analyses; used throughout (Lemmas 2.5–2.7, 3.3, etc.).
  • domain assumption Problem (1.1) is solvable (Assumption 2.3)
    Existence of x* — standard.
  • domain assumption Global boundedness of the Hessian, ||∇²f(x)|| ≤ κ_H (Assumption 2.4)
    Restrictive global assumption, needed only for Alg 1: Lemmas 3.6, 3.11, 3.18 (bounded gradient at λ_k = 0 iterates and bisection complexity). Not stated in the abstract's rate claim.
  • domain assumption Non-degenerate minimizer, ∇²f(x*) ⪰ νI (Assumption 3.13)
    Required for the quadratic local-convergence claim (Lemmas 3.14–3.19, Theorem 3.20); the abstract's 'quadratic convergence' is conditional on this.
  • standard math Estimating-sequence framework and its standard lemmas (Nesterov)
    Uniform convexity bounds (3.5)-(3.6), min of cubic-plus-linear, and growth bound (3.4) come from Nesterov [44]; the min constant is where the factor-8 error occurs.
  • standard math KKT optimality conditions for the (TR+) subproblem (Lemma 2.8)
    Imported from Conn et al. [12]; underpins the stability and bisection analyses.
  • standard math Quadratic convergence of Newton's method near non-degenerate minimizers (Lemma 3.14)
    Imported from Nesterov [44, Thm 1.2.5]; the 'quadratic local convergence' of Alg 1 is literally Newton's theorem applied inside the diving track.
  • standard math Monteiro–Svaiter framework lemmas (curve y(σ), growth of A_k)
    Alg 2's iteration complexity inherits MS 2013 / Carmon et al. 2022 results (Lemma 4.3, Lemma C.3).
  • ad hoc to paper Constant balance in Lemma 3.4's estimate-sequence induction
    The proof implicitly assumes the cubic-minimum constant (1/(3√3√M)) matches the descent bound (√6/6)/√M; the exact constant is 8× larger, so the assumption fails. This is the load-bearing defective premise.
invented entities (2)
  • (TR+) oracle — trust-region subproblem with a primal shift σ
    purpose: Steers the dual multiplier λ into a stable window and supplies λ as a free local-geometry signal
    Algorithmic gadget (not a physical entity). Its key property — stability M||d|| ≤ σ+λ ≤ 2M||d|| — is proven in-paper (Lemma 3.3, Section 2) and reused from the authors' own [32]; there is no outside-paper falsifiable handle.
  • Local Detection (LD) subroutine with tracks 'local diving' and 'R&B'
    purpose: Switch into Newton steps when λ = 0 signals the quadratic region; preserve global acceleration when the signal is false
    New algorithmic machinery introduced in this paper; its soundness is exactly what Sections 3.2–3.3 argue, so its evidence is in-paper only.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Accelerating Trust-Region Methods: An Attempt to Balance Global and Local Efficiency." pith.science (2026). https://pith.science/paper/BCZZ4YBF

@misc{pith2026251100680,
  author       = {Pith},
  title        = {Pith review of: Accelerating Trust-Region Methods: An Attempt to Balance Global and Local Efficiency},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BCZZ4YBF}},
  note         = {Machine review of arXiv:2511.00680}
}
read the original abstract

Balancing global efficiency and local convergence remains a central challenge in second-order methods for unconstrained convex optimization problems. Newton's method enjoys fast local convergence but may diverge when initialized far from the solution. In contrast, accelerated second-order methods provide global guarantees but typically suffer from slower local convergence. This raises the fundamental question of to what extent global acceleration can be achieved without sacrificing strong local convergence. In this paper, we tackle this challenge by proposing the first accelerated trust-region-type methods and leveraging their inherent primal-dual information. Our primary contribution is the Accelerated Trust-Region method with Local Detection, which utilizes the Lagrange multiplier to detect local regions and achieves a global oracle complexity of \tilde{O}(\epsilon^{-1/3}), while maintaining quadratic local convergence. We further examine the trade-off that arises when global convergence is pushed to the limit. Specifically, we introduce the Accelerated Trust-Region Extragradient Method, which achieves a global oracle complexity of \tilde{O}(\epsilon^{-2/7}) but no longer enjoys quadratic local convergence. This reveals a phase-transition-like phenomenon in accelerated trust-region-type methods: quadratic local convergence is preserved under moderate global acceleration, but it breaks down when pursuing extreme global efficiency. Numerical experiments are consistent with the theoretical predictions and illustrate the global-local trade-off.

Figures

Figures reproduced from arXiv: 2511.00680 by the authors.

Figure 1
Figure 1. Logistic regression using the LIBSVM datasets [PITH_FULL_IMAGE:figures/full_fig_p028_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. On the Universality of Simple Trust-Region Algorithms

    math.OC 2026-07 accept novelty 6.0 of 10

    Classical and modified-ratio trust-region methods reach the optimal O(ε^{-1/(1+ν)}) convex and O(ε^{-(2+ν)/(1+ν)}) nonconvex complexity for any Hölder ν∈[0,1] without knowing ν.

Reference graph

Works this paper leans on

63 extracted references · 5 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Eigenvalue-based algorithm and analysis for nonconvex QCQP with one constraint.Mathematical Programming, 173(1-2):79–116, 2019

    Satoru Adachi and Yuji Nakatsukasa. Eigenvalue-based algorithm and analysis for nonconvex QCQP with one constraint.Mathematical Programming, 173(1-2):79–116, 2019

  2. [2]

    Solving the trust- region subproblem by a generalized eigenvalue problem.SIAM Journal on Optimiza- tion, 27(1):269–291, January 2017

    Satoru Adachi, Satoru Iwata, Yuji Nakatsukasa, and Akiko Takeda. Solving the trust- region subproblem by a generalized eigenvalue problem.SIAM Journal on Optimiza- tion, 27(1):269–291, January 2017

  3. [3]

    Inexact tensor methods and their application to stochastic convex opti- mization.Optimization Methods and Software, pages 1–42, 2023

    Artem Agafonov, Dmitry Kamzolov, Pavel Dvurechensky, Alexander Gasnikov, and Martin Tak´ aˇ c. Inexact tensor methods and their application to stochastic convex opti- mization.Optimization Methods and Software, pages 1–42, 2023

  4. [4]

    Advancing the lower bounds: an ac- celerated, stochastic, second-order method with optimal adaptation to inexactness

    Artem Agafonov, Dmitry Kamzolov, Alexander Gasnikov, Ali Kavis, Kimon Anton- akopoulos, Volkan Cevher, and Martin Tak´ aˇ c. Advancing the lower bounds: an ac- celerated, stochastic, second-order method with optimal adaptation to inexactness. In The Twelfth International Conference on Learning Representations, 2024

  5. [5]

    Estimate sequence methods: extensions and approximations.Institute for Operations Research, ETH, Z¨ urich, Switzerland, 2(1), 2009

    Michel Baes. Estimate sequence methods: extensions and approximations.Institute for Operations Research, ETH, Z¨ urich, Switzerland, 2(1), 2009

  6. [6]

    Knitro: An integrated package for nonlinear optimization.Large-scale nonlinear optimization, pages 35–59, 2006

    Richard H Byrd, Jorge Nocedal, and Richard A Waltz. Knitro: An integrated package for nonlinear optimization.Large-scale nonlinear optimization, pages 35–59, 2006

  7. [7]

    Op- timal and adaptive monteiro-svaiter acceleration.Advances in Neural Information Pro- cessing Systems, 35:20338–20350, 2022

    Yair Carmon, Danielle Hausler, Arun Jambulapati, Yujia Jin, and Aaron Sidford. Op- timal and adaptive monteiro-svaiter acceleration.Advances in Neural Information Pro- cessing Systems, 35:20338–20350, 2022

  8. [8]

    Coralia Cartis, Nicholas IM Gould, and Ph L Toint. On the complexity of steepest descent, newton’s and regularized newton’s methods for nonconvex unconstrained op- timization problems.Siam journal on optimization, 20(6):2833–2852, 2010

Show all 63 references
  1. [9]

    Adaptive cubic regularisation methods for unconstrained optimization

    Coralia Cartis, Nicholas IM Gould, and Philippe L Toint. Adaptive cubic regularisation methods for unconstrained optimization. part i: motivation, convergence and numerical results.Mathematical Programming, 127(2):245–295, 2011

  2. [10]

    SIAM, 2022

    Coralia Cartis, Nicholas IM Gould, and Philippe L Toint.Evaluation Complexity of Al- gorithms for Nonconvex Optimization: Theory, Computation and Perspectives. SIAM, 2022

  3. [11]

    Accelerating adaptive cubic regularization of newton’s method via random sampling.The Journal of Machine Learning Research, 23(1):3904–3941, 2022

    Xi Chen, Bo Jiang, Tianyi Lin, and Shuzhong Zhang. Accelerating adaptive cubic regularization of newton’s method via random sampling.The Journal of Machine Learning Research, 23(1):3904–3941, 2022. 29

  4. [12]

    SIAM, 2000

    Andrew R Conn, Nicholas IM Gould, and Philippe L Toint.Trust Region Methods. SIAM, 2000

  5. [13]

    A trust region algo- rithm with a worst-case iteration complexity ofO(ϵ −3/2) for nonconvex optimization

    Frank E Curtis, Daniel P Robinson, and Mohammadreza Samadi. A trust region algo- rithm with a worst-case iteration complexity ofO(ϵ −3/2) for nonconvex optimization. Mathematical Programming, 162:1–32, 2017

  6. [14]

    Concise complexity analyses for trust region methods.Optimization Letters, 12:1713–1724, 2018

    Frank E Curtis, Zachary Lubberts, and Daniel P Robinson. Concise complexity analyses for trust region methods.Optimization Letters, 12:1713–1724, 2018

  7. [15]

    Trust- region newton-cg with strong second-order complexity guarantees for nonconvex opti- mization.SIAM Journal on Optimization, 31(1):518–544, 2021

    Frank E Curtis, Daniel P Robinson, Cl´ ement W Royer, and Stephen J Wright. Trust- region newton-cg with strong second-order complexity guarantees for nonconvex opti- mization.SIAM Journal on Optimization, 31(1):518–544, 2021

  8. [16]

    SIAM, 1996

    John E Dennis Jr and Robert B Schnabel.Numerical methods for unconstrained opti- mization and nonlinear equations. SIAM, 1996

  9. [17]

    Contracting proximal methods for smooth convex optimization.SIAM Journal on Optimization, 30(4):3146–3169, 2020

    Nikita Doikov and Yurii Nesterov. Contracting proximal methods for smooth convex optimization.SIAM Journal on Optimization, 30(4):3146–3169, 2020

  10. [18]

    Cardinal Optimizer (COPT) User Guide, October 2022

    Dongdong Ge, Qi Huangfu, Zizhuo Wang, Jian Wu, and Yinyu Ye. Cardinal Optimizer (COPT) User Guide, October 2022

  11. [19]

    Accelerated gradient methods for nonconvex non- linear and stochastic programming.Mathematical Programming, 156(1):59–99, 2016

    Saeed Ghadimi and Guanghui Lan. Accelerated gradient methods for nonconvex non- linear and stochastic programming.Mathematical Programming, 156(1):59–99, 2016

  12. [20]

    On the convergence and worst-case complexity of trust-region and regularization methods for unconstrained optimization.Mathematical Programming, 152(1):491–520, 2015

    Geovani N Grapiglia, Jinyun Yuan, and Ya-xiang Yuan. On the convergence and worst-case complexity of trust-region and regularization methods for unconstrained optimization.Mathematical Programming, 152(1):491–520, 2015

  13. [21]

    A consistently adaptive trust-region method.Advances in Neural Information Processing Systems, 35:6640–6653, 2022

    Fadi Hamad and Oliver Hinder. A consistently adaptive trust-region method.Advances in Neural Information Processing Systems, 35:6640–6653, 2022

  14. [22]

    A simple and practical adaptive trust-region method

    Fadi Hamad and Oliver Hinder. A simple and practical adaptive trust-region method. arXiv preprint arXiv:2412.02079, 2024

  15. [23]

    Ho- mogeneous second-order descent framework: a fast alternative to Newton-type meth- ods.Mathematical Programming, May 2025

    Chang He, Yuntian Jiang, Chuwen Zhang, Dongdong Ge, Bo Jiang, and Yinyu Ye. Ho- mogeneous second-order descent framework: a fast alternative to Newton-type meth- ods.Mathematical Programming, May 2025. ISSN 1436-4646

  16. [24]

    A second-order cone based approach for solving the trust-region subproblem and its variants.SIAM Journal on Optimization, 27(3):1485–1512, 2017

    Nam Ho-Nguyen and Fatma Kilinc-Karzan. A second-order cone based approach for solving the trust-region subproblem and its variants.SIAM Journal on Optimization, 27(3):1485–1512, 2017

  17. [25]

    An approximation-based regularized extra-gradient method for monotone variational inequalities.SIAM Journal on Optimization, 35(3): 1469–1497, 2025

    Kevin Huang and Shuzhong Zhang. An approximation-based regularized extra-gradient method for monotone variational inequalities.SIAM Journal on Optimization, 35(3): 1469–1497, 2025. 30

  18. [26]

    Inexact and implementable acceler- ated newton proximal extragradient method for convex optimization.arXiv preprint arXiv:2402.11951, 2024

    Ziyu Huang, Bo Jiang, and Yuntian Jiang. Inexact and implementable acceler- ated newton proximal extragradient method for convex optimization.arXiv preprint arXiv:2402.11951, 2024

  19. [27]

    A unified adaptive tensor approximation scheme to accelerate composite convex optimization.SIAM Journal on Optimization, 30(4):2897–2926, 2020

    Bo Jiang, Tianyi Lin, and Shuzhong Zhang. A unified adaptive tensor approximation scheme to accelerate composite convex optimization.SIAM Journal on Optimization, 30(4):2897–2926, 2020

  20. [28]

    An optimal high-order tensor method for convex optimization.Mathematics of Operations Research, 46(4):1390–1412, 2021

    Bo Jiang, Haoyue Wang, and Shuzhong Zhang. An optimal high-order tensor method for convex optimization.Mathematics of Operations Research, 46(4):1390–1412, 2021

  21. [29]

    Generalized optimistic methods for convex-concave saddle point problems.arXiv preprint arXiv:2202.09674, 2022

    Ruichen Jiang and Aryan Mokhtari. Generalized optimistic methods for convex-concave saddle point problems.arXiv preprint arXiv:2202.09674, 2022

  22. [30]

    Accelerated quasi-newton proximal extragradient: Faster rate for smooth convex optimization.Advances in Neural Information Processing Systems, 36, 2024

    Ruichen Jiang and Aryan Mokhtari. Accelerated quasi-newton proximal extragradient: Faster rate for smooth convex optimization.Advances in Neural Information Processing Systems, 36, 2024

  23. [31]

    H¨ olderian Error Bounds and Kurdyka- Lojasiewicz In- equality for the Trust Region Subproblem.Mathematics of Operations Research, 47(4): 3025–3050, November 2022

    Rujun Jiang and Xudong Li. H¨ olderian Error Bounds and Kurdyka- Lojasiewicz In- equality for the Trust Region Subproblem.Mathematics of Operations Research, 47(4): 3025–3050, November 2022. ISSN 0364-765X. doi: 10.1287/moor.2021.1243

  24. [32]

    Be- yond nonconvexity: A universal trust-region method with new analyses.arXiv preprint arXiv:2311.11489, 2024

    Yuntian Jiang, Chang He, Chuwen Zhang, Dongdong Ge, Bo Jiang, and Yinyu Ye. Be- yond nonconvexity: A universal trust-region method with new analyses.arXiv preprint arXiv:2311.11489, 2024

  25. [33]

    The first optimal acceleration of high-order methods in smooth convex optimization.Advances in Neural Information Processing Systems, 35:35339–35351, 2022

    Dmitry Kovalev and Alexander Gasnikov. The first optimal acceleration of high-order methods in smooth convex optimization.Advances in Neural Information Processing Systems, 35:35339–35351, 2022

  26. [34]

    An optimal method for stochastic composite optimization.Mathemat- ical Programming, 133(1):365–397, 2012

    Guanghui Lan. An optimal method for stochastic composite optimization.Mathemat- ical Programming, 133(1):365–397, 2012

  27. [35]

    Perseus: A simple and optimal high-order method for variational inequalities.Mathematical Programming, 209(1):609–650, 2025

    Tianyi Lin and Michael I Jordan. Perseus: A simple and optimal high-order method for variational inequalities.Mathematical Programming, 209(1):609–650, 2025

  28. [36]

    Explicit second-order min-max optimization methods with optimal convergence guarantee.arXiv preprint arXiv:2210.12860, 2022

    Tianyi Lin, Panayotis Mertikopoulos, and Michael I Jordan. Explicit second-order min-max optimization methods with optimal convergence guarantee.arXiv preprint arXiv:2210.12860, 2022

  29. [37]

    M Marques Alves. Variants of the a-hpe and large-step a-hpe algorithms for strongly convex problems with applications to accelerated high-order tensor methods.Optimiza- tion Methods and Software, 37(6):2021–2051, 2022

  30. [38]

    Regularized Newton Method with GlobalO(1/k 2) Conver- gence.SIAM Journal on Optimization, 33(3):1440–1462, 2023

    Konstantin Mishchenko. Regularized Newton Method with GlobalO(1/k 2) Conver- gence.SIAM Journal on Optimization, 33(3):1440–1462, 2023. 31

  31. [39]

    An accelerated hybrid proximal extra- gradient method for convex optimization and its implications to second-order methods

    Renato DC Monteiro and Benar Fux Svaiter. An accelerated hybrid proximal extra- gradient method for convex optimization and its implications to second-order methods. SIAM Journal on Optimization, 23(2):1092–1125, 2013

  32. [40]

    J. J. Mor´ e.Recent Developments in Algorithms and Software for Trust Region Methods, pages 258–287. Springer Berlin Heidelberg, Berlin, Heidelberg, 1983

  33. [41]

    Newton’s method

    Jorge J Mor´ e and Danny C Sorensen. Newton’s method. Technical report, Argonne National Lab.(ANL), Argonne, IL (United States), 1982

  34. [42]

    Accelerating the cubic regularization of newton’s method on convex problems.Mathematical Programming, 112(1):159–181, 2008

    Yu Nesterov. Accelerating the cubic regularization of newton’s method on convex problems.Mathematical Programming, 112(1):159–181, 2008

  35. [43]

    A method for solving the convex programming problem with conver- gence rateO(1/k 2)

    Yurii Nesterov. A method for solving the convex programming problem with conver- gence rateO(1/k 2). InDokl. Akad. Nauk. SSSR, volume 269, page 543, 1983

  36. [44]

    Springer, 2018

    Yurii Nesterov.Lectures on Convex Optimization, volume 137. Springer, 2018

  37. [45]

    Implementable tensor methods in unconstrained convex optimization

    Yurii Nesterov. Implementable tensor methods in unconstrained convex optimization. Mathematical Programming, 186:157–183, 2021

  38. [46]

    Cubic regularization of newton method and its global performance.Mathematical Programming, 108(1):177–205, 2006

    Yurii Nesterov and Boris T Polyak. Cubic regularization of newton method and its global performance.Mathematical Programming, 108(1):177–205, 2006

  39. [47]

    Springer, 1999

    Jorge Nocedal and Stephen J Wright.Numerical Optimization. Springer, 1999

  40. [48]

    Ten- sor methods for strongly convex strongly concave saddle point problems and strongly monotone variational inequalities.arXiv preprint arXiv:2012.15595, 2020

    Petr Ostroukhov, Rinat Kamalov, Pavel Dvurechensky, and Alexander Gasnikov. Ten- sor methods for strongly convex strongly concave saddle point problems and strongly monotone variational inequalities.arXiv preprint arXiv:2012.15595, 2020

  41. [49]

    Newton’s method and its use in optimization.European Journal of Operational Research, 181(3):1086–1096, 2007

    Boris T Polyak. Newton’s method and its use in optimization.European Journal of Operational Research, 181(3):1086–1096, 2007

  42. [50]

    PDFO: a cross-platform package for powell’s derivative-free optimization solvers.Mathematical Programming Computation, pages 1–25, 2024

    Tom M Ragonneau and Zaikun Zhang. PDFO: a cross-platform package for powell’s derivative-free optimization solvers.Mathematical Programming Computation, pages 1–25, 2024

  43. [51]

    Santos, and Danny C

    Marielba Rojas, Sandra A. Santos, and Danny C. Sorensen. A new matrix-free algo- rithm for the large-scale trust-region subproblem.SIAM Journal on Optimization, 11 (3):611–646, January 2001

  44. [52]

    Trust region policy optimization

    John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. InInternational conference on machine learning, pages 1889–1897. PMLR, 2015. 32

  45. [53]

    Unified acceleration of high-order algorithms under general H¨ older continuity.SIAM Journal on Optimization, 31(3):1797–1826, January 2021

    Chaobing Song, Yong Jiang, and Yi Ma. Unified acceleration of high-order algorithms under general H¨ older continuity.SIAM Journal on Optimization, 31(3):1797–1826, January 2021

  46. [54]

    Vavasis and Richard Zippel

    Stephen A. Vavasis and Richard Zippel. Proving polynomial-time for sphere-constrained quadratic programming. Technical report, Cornell University, 1990

  47. [55]

    The generalized trust region subproblem: solution complexity and convex hull results.Mathematical Programming, 191(2):445– 486, 2022

    Alex L Wang and Fatma Kılın¸ c-Karzan. The generalized trust region subproblem: solution complexity and convex hull results.Mathematical Programming, 191(2):445– 486, 2022

  48. [56]

    Accelerated first-order primal-dual proximal methods for linearly con- strained composite convex programming.SIAM Journal on Optimization, 27(3):1459– 1484, 2017

    Yangyang Xu. Accelerated first-order primal-dual proximal methods for linearly con- strained composite convex programming.SIAM Journal on Optimization, 27(3):1459– 1484, 2017

  49. [57]

    Accelerated primal–dual proximal block coordinate updating methods for constrained convex optimization.Computational Optimization and Applications, 70(1):91–128, 2018

    Yangyang Xu and Shuzhong Zhang. Accelerated primal–dual proximal block coordinate updating methods for constrained convex optimization.Computational Optimization and Applications, 70(1):91–128, 2018

  50. [58]

    Trust region based adversarial attack on neural networks

    Zhewei Yao, Amir Gholami, Peng Xu, Kurt Keutzer, and Michael W Mahoney. Trust region based adversarial attack on neural networks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11350–11359, 2019

  51. [59]

    A New Complexity Result on Minimization of a Quadratic Function with a Sphere Constraint

    Yinyu Ye. A New Complexity Result on Minimization of a Quadratic Function with a Sphere Constraint. InRecent Advances in Global Optimization, volume 176, pages 19–31. Princeton University Press, 1991

  52. [60]

    Second Order Optimization Algorithms I, 2005

    Yinyu Ye. Second Order Optimization Algorithms I, 2005. URLhttps://web. stanford.edu/class/msande311/lecture12.pdf

  53. [61]

    A review of trust region algorithms for optimization

    Ya-Xiang Yuan. A review of trust region algorithms for optimization. InIciam, vol- ume 99, pages 271–282, 2000

  54. [62]

    Recent advances in trust region algorithms.Mathematical Program- ming, 151:249–281, 2015

    Ya-Xiang Yuan. Recent advances in trust region algorithms.Mathematical Program- ming, 151:249–281, 2015

  55. [63]

    1 σ+ ∥y(σ−)−y(σ +)∥+ M σ2 + ∥y(σ−)−y(σ +)∥2 + 2M σ+ ∥y(σ−)−y(σ +)∥ ·ψ+ # +

    Chuwen Zhang, Chang He, Yuntian Jiang, Chenyu Xue, Bo Jiang, Dongdong Ge, and Yinyu Ye. A homogeneous second-order descent method for nonconvex optimization. Mathematics of Operations Research, May 2025. ISSN 0364-765X. 33 A Technical proofs in Section 2 Proof to Theorem 2.6 P...

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.